Kimi K3 three months later: 1.56TB is too big to run locally and rivals have caught up
Sanbi_Ai · reddit · 2026-10-17
A retrospective on Kimi K3 after the hype: the full 2.8T-param model weighs 1.56TB, Unsloth's smallest recommended quant is 594GB (610GB RAM+VRAM), 1-bit agrees with the full model on only 79% of next tokens, 20 tok/s on B200s and 43 tok/s via Moonshot's own API ($10+/task on AA agentic evals). Its AA index has slipped to 44, now tied by Step 5 Preview and nearly matched by GLM-5.3 (45). The license isn't MIT: companies with >$20M revenue over any 12 months need a separate deal with Moonshot for commercial use, even if K3 itself earns nothing.
More from Infra
- Epoch AI chart: xAI owns fewer GPUs than CoreWeave, far behind hyperscalers — bookwormengr · 2026-10-17
- Nvidia CEO Jensen Huang says AI stock selloff is temporary: "All of this is going to pass" — Polymarket · 2026-10-17
- Cloudflare rolls out full ZDR on Workers AI: no data retention, no training on your data — michellechen · 2026-10-17
- Modal explains its speed moat: ImageFS tiered filesystem now tricking VM ext4 — dejavucoder · 2026-10-17
- AWS built a production eval framework on Bedrock AgentCore: agents fail silently — AWS ML Blog · 2026-10-17
- Why are you running agent fleets on a 16GB laptop? Reddit post warns of frozen machines and dead batteries — ConceptNext5110 · 2026-10-17